Study Guide

Copilot is an orchestration engine that coordinates three things: large language models, content in Microsoft Graph that you have permission to access, and the Microsoft 365 apps you work in every day. Everything it does happens inside the Microsoft 365 service boundary — your tenant — where your existing security, compliance, and privacy policies already apply. Operating inside that boundary does not give Copilot tenant-wide visibility: data access is always scoped to the signed-in user's permissions.

The single most tested commitment is that prompts, responses, and data accessed through Microsoft Graph aren't used to train the foundation LLMs. The second is permission trimming: Copilot only surfaces organizational data the individual user has at least view permissions for. The semantic index honors that same identity-based access boundary, so grounding only ever reaches content the current user is authorized to see. If a scenario says a user can't find a file through Copilot, the answer is almost always a permissions or labeling problem, not a Copilot problem.

Your interactions are stored. The user's prompt and Copilot's response — together called the content of interactions — become the user's Copilot activity history. That data is encrypted at rest, is processed and stored under the same contractual commitments as the rest of your Microsoft 365 content, and is available to admins through Content search and Microsoft Purview, including retention policies and eDiscovery. Users can delete their own activity history from the My Account portal.

Web search is the one place data leaves the service boundary, and even then only a little. When web search is on, Copilot parses your prompt, identifies terms where the web would help, and generates a short search query — a few words, not your whole prompt — that it sends to the Bing search service with user and tenant identifiers removed. Entire files, entire prompts, and anything identifying from Microsoft Entra ID are not included. Those generated queries aren't used to improve Bing, aren't shared with advertisers, and aren't used to train foundation models.

Enterprise data protection (EDP) is the umbrella term for these commitments under the Data Protection Addendum and Product Terms, and it covers both Microsoft Copilot and Microsoft Copilot Chat. Note two carve-outs that make good exam distractors: the EU Data Boundary and HIPAA compliance do not apply to generated web search queries.

When you submit a prompt, Copilot doesn't hand it straight to the model. It preprocesses the prompt by grounding it: accessing Microsoft Graph in your tenant, gathering the relevant emails, chats, documents, meetings, and calendar entries you have access to, and combining them with your working context. The grounded prompt goes to the LLM, and the response comes back to the app. Grounding is what makes an answer specific to your organization instead of generic.

Work grounding is the licensed capability. Microsoft Copilot with a license reaches your work data through Microsoft Graph. Microsoft Copilot Chat without that license is web-grounded: it uses the web and whatever you paste or upload, but it can't reason over your emails, meetings, and SharePoint content. This distinction drives a large share of exam questions — whenever a scenario says "summarize my meetings from last week" or "find the decision my colleague made," work grounding and therefore a license is required.

Web grounding is controlled at two levels. Admins use the Allow web search in Copilot policy in the Cloud Policy service for Microsoft 365, which can enable web search, disable it, or disable it in Work mode while leaving it on in Web mode and Copilot Chat. Users have a Web content toggle; if the admin has turned web search off, the toggle is dimmed and unusable. Web search query citations — showing the exact queries sent to Bing — appear only in Copilot Chat, not in the Copilot pane inside Word or PowerPoint, and only for 24 hours.

In-app context matters too. Copilot in Word behaves differently from Copilot Chat because the document you have open is part of the context. In Word and PowerPoint you steer the source through the prompt itself — reference a file with /filename to draw from your organizational content, or describe a web source to draw from the web. Copilot in Outlook uses an explicit Work and Web toggle instead.

In a chat experience you are in the loop. You write a prompt, read the result, decide what to ask next, and refine as you go. Copilot Chat is general-purpose: drafting, summarizing, researching, analyzing, brainstorming. Every turn is your decision, and the conversation is open-ended.

An agent is different in kind, not just in degree. It's a specialized assistant configured ahead of time for a particular type of work — an HR agent that answers benefits questions, a project agent grounded in one set of files, a sales agent for account workflows. You reach for an agent when you need consistent, specialized help for a repeated task: always the same response format, or always the same knowledge source. For one-off questions and general research, ordinary Copilot Chat works just as well.

An agent is made of a few configurable parts, and knowing them is worth marks: a name, description, and icon that identify it; instructions that define its persona, tasks, and response style; knowledge sources that ground its answers; capabilities such as code interpreter and image generator; and suggested (starter) prompts that show users what it's for. The description is what helps Copilot decide when to invoke the agent; the instructions are what shape how it behaves once invoked.

Agents appear in the Agent Store, a curated marketplace of prebuilt agents from Microsoft, partners, and internal teams. Microsoft-built examples include Writing Coach, Prompt Coach, and Visual Creator. The advanced reasoning agents Researcher and Analyst are also in the store but require a Microsoft 365 Copilot license.

The signal for building your own agent is repetition with constraints. If several people keep asking the same category of question, if the answer should always come from a particular SharePoint site, or if the output should always follow one format, an agent captures that once instead of relying on everyone writing a good prompt every time. Prebuilt agents from the Agent Store cover many common tasks, so check there first — building your own is for your team's specific processes, knowledge sources, tone, and format.

Agent Builder is the lightweight, no-code path inside Microsoft 365 Copilot. You describe the agent in natural language and Copilot configures it for you, or you skip to the Configure tab and set each field yourself. It's designed for business users and produces declarative agents.

Copilot Studio is the more capable authoring environment. Reach for it when you need what Agent Builder can't do: publishing an agent across multiple channels, or strict control over knowledge. Agent Builder's Only use specified sources setting prioritizes your sources but cannot fully block the model's general knowledge; Copilot Studio can. That trade-off is a favourite exam distinction.

Note also that agent creation has to be enabled by your administrator, and that agents which reason over work data — emails, meetings, SharePoint, CRM records through Copilot connectors — require either metered consumption billing or a Microsoft 365 Copilot license.

Copilot in Word drafts new content from a prompt, rewrites selected text for tone or length, summarizes or answers questions about an open document in Chat only mode, converts text into tables, and generates images inline. Reference a specific file, email, or meeting by typing / in the Copilot pane.

Copilot in PowerPoint creates a presentation from a prompt or from an existing document — reference the file with / and you can draw on up to five files at once, mixing a Word document, an Excel file, and a PDF in one prompt. It also generates slide images, applies a new design across the whole deck in one step, summarizes a deck, and adds or expands individual slides.

Copilot in Excel works inside the spreadsheet: exploring, charting, and transforming data that's already there. For deeper analysis across several files with citations and a shareable report, the Analyst agent is the right tool instead.

Copilot in Outlook drafts complete emails and meeting requests from a single prompt, and offers coaching on an existing draft through Get Coaching, which suggests improvements to tone, clarity, and reader sentiment. Outlook is also where you'll see the explicit Work and Web toggle, and where switching from Default mode to Allow actions lets Copilot take real actions such as creating a calendar event after you approve it.

Across all of these, the licensing line is the same: Copilot Chat alone is web-grounded and free-tier, while work-data grounding, Researcher, Analyst, and scheduled prompts require a Microsoft 365 Copilot license. Expect the exam to test this boundary more often than any individual app feature.

Inaccuracies — also called fabrications or hallucinations — are responses that sound confident but contain incorrect facts, numbers, or citations. Generative output is not guaranteed to be factual. The mitigation is not to avoid Copilot but to treat every response as a draft you verify.

Prompt injection is malicious instruction hidden inside user input or external content that tricks Copilot into ignoring its rules or taking unintended actions. Microsoft applies jailbreak and cross-prompt injection attack (XPIA) classifiers that analyze inputs and block high-risk prompts before the model runs, but platform safeguards don't remove your responsibility to be careful with untrusted content. Note the related control: a DLP policy can exclude email received from external senders from grounding specifically to reduce prompt-injection exposure.

Over-reliance is treating Copilot's output as final without applying your own judgment, which lets errors and biases reach your work unchecked. It's a human risk rather than a technical one, and the exam treats it as seriously as the other two.

Sitting behind all of this are Microsoft's six responsible AI principles: fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. Copilot puts them into practice through grounded responses with citations, content moderation through classifiers and filters, and privacy by design.

Microsoft Learn sets out a sequence: review the full response for relevance, completeness, and audience fit; check the sources by opening each citation and confirming it exists, is trustworthy, and actually supports the claim; verify key details such as facts, numbers, dates, names, and quotations against trusted sources; check for sensitive information that shouldn't be shared; get expert review for specialized content you can't verify yourself; and approve the final version, taking responsibility for it.

The exam's real test is proportionality. Verification should match the stakes. A casual internal summary doesn't need a subject-matter expert; a customer-facing financial claim does. When a question offers several plausible verification steps, pick the one that matches the risk described in the scenario rather than the most thorough one available.

Human review is the named mitigation for high-stakes work. Copilot's own design reflects this — Microsoft describes its capabilities as providing drafts and summaries that give you a chance to review generated content rather than fully automating the task.

Before you type, check whether the information is sensitive: customer records, financial details, health information, credentials, or trade secrets. The habits Learn recommends are concrete. Share only what the task needs — paste a summary or excerpt rather than an entire document or dataset. Respect sensitivity labels: files labeled confidential or highly confidential carry restrictions Copilot honors, but you remain responsible for handling that content appropriately. Review responses before sharing them externally, confirming nothing has crept in that shouldn't leave the organization.

Oversharing is the organizational version of the same risk. Because Copilot surfaces everything a user can already open, over-permissive SharePoint sites and broad sharing links suddenly become very visible. The fix is on the permissions side, not the Copilot side: tighten access, apply labels, and use Microsoft Purview Data Security Posture Management (DSPM) for AI to run data risk assessments that identify oversharing before it becomes a problem.

One habit is easy to overlook and gets tested: report unexpected access. If Copilot surfaces content you don't believe you should be able to see, tell IT or security so they can check permissions and sharing settings. It's a signal that something is misconfigured.

Sensitivity labels are the first layer. When a label applies encryption, a user needs the EXTRACT usage right as well as VIEW before Copilot will return that data. In Copilot Chat, responses display the highest-priority label from the data used, so users see how sensitive the underlying material was.

Label inheritance is heavily tested. When you use Copilot in Word, PowerPoint, or Outlook to create new content based on a labeled item, the new content automatically inherits the source file's sensitivity label and its protection settings. If multiple files were used, the highest-priority label wins. The user can still override an inherited label unless mandatory labeling is in force.

DLP adds a second, policy-driven layer through the Microsoft 365 Copilot and Copilot Chat policy location, which is available only in the Custom policy template. It supports four combinations worth knowing. Content contains sensitive information types with the action Processing prompts stops Copilot responding at all when a prompt contains, say, a credit card or passport number. The same condition with Performing Web Searches lets the response proceed but blocks the web as a grounding source. Content contains sensitivity labels excludes labeled files and emails from being processed — importantly, the item can still appear in the response's citations, but its content isn't used. And Email is received from external users (preview) excludes external email from grounding, summarization, and citation, specifically to reduce prompt-injection risk.

Two limits are easy to miss. You can't combine the sensitive-info-types condition and the sensitivity-labels condition in the same rule — use separate rules in the same policy. And DLP can't scan the contents of files uploaded directly into a prompt; it only inspects the text you type.

The four elements are worth memorizing in order, because questions often present a weak prompt and ask what's missing. Context is why you need it and who's involved. Goal is what you want Copilot to do and in what form — bullet points, an essay, an email, a draft. Source is where Copilot should look, whether that's an uploaded file, a referenced document, or a type of web source. Expectations is how the answer should be delivered: the audience, the tone, the level of detail.

A clear goal is the foundation, but adding context, sources, and expectations is what moves a response from generic to usable. Copilot responds best to detailed prompts — unlike traditional search, more words generally help rather than hurt.

Iteration is the second half of the skill. Rather than accepting the first response, send follow-up prompts that clarify what was missing: make the tone more formal, cut this to two sentences, add a section on risks. Copilot is designed for back-and-forth refinement, and questions that describe a user dissatisfied with a first response are usually testing whether you reach for iteration rather than starting over.

Two smaller features round this out. Voice input is available through Dictation in the prompt box and Read Aloud for responses, though real-time two-way voice chat requires a Microsoft 365 Copilot license.

Type / followed by a file or folder name to link directly to it — for example, "Summarize the key points from /Q4MarketingStrategy.pptx". This is the workhorse reference and appears in both Copilot Chat and the Copilot pane inside Word and PowerPoint. The same / syntax references a meeting by name when you want a follow-up grounded in what was actually said.

Type # followed by a topic, meeting name, or recurring event to reference it — for example, "What were the action items from #WeeklyOpsSync?". This is a licensed capability, because it reaches work content. Type @ followed by a person's name to mention them, which also brings their perspective into collaborative surfaces such as a shared Copilot Page.

Choosing between work and web sources is a judgment question the exam likes. If the scenario needs internal facts — what a colleague decided, what's in last quarter's report, what a meeting concluded — the answer must involve work grounding and therefore a license. If it needs market data, competitor information, or industry trends, web grounding is right. Some prompts want both: Copilot can return a response split into a "From your company's data" section and a "From the web" section.

Referencing work content such as meetings and chats requires a Microsoft 365 Copilot license. That single sentence resolves a surprising number of exam questions.

Prompt Lab (previously called Prompt Gallery) is the catalog of prompts in Copilot Chat. Open it by selecting the prompt pills or the ... below the chat input box. It shows Microsoft-curated prompts designed to work across roles and industries, the prompts you've saved yourself, prompts shared with teams you belong to, and collections published by your organization's admin using your company's own workflows and terminology.

To save a prompt: enter and run it in Copilot Chat, look for the bookmark icon next to your prompt after Copilot responds (on some devices you need to hover over the prompt text to reveal it), select it, and enter a name. It then appears under the Your prompts tab.

One limitation is worth remembering because it makes a clean exam question: you can't edit a saved prompt in place. To update one, run or copy it from Prompt Lab, make your changes in the chat input, and save it as a new prompt — then delete the original to avoid duplicates.

Know the three tabs and what feeds each: Suggested comes from Microsoft, Your prompts comes from what you bookmarked, and Team prompts comes from what someone shared with a team you're on.

Scheduling lets Copilot run a prompt automatically at a set time — summarizing your inbox every morning at 8 AM, for example — without you being online. The steps are: type and run the prompt you want to automate, look for the scheduling option next to the prompt once the response appears (hover to reveal it on some devices), select Schedule this prompt, choose the frequency such as daily at 8 AM or weekly on Fridays, and select Save. You can create up to 10 scheduled prompts, and scheduled runs are marked in your conversation history.

Managing them is a separate flow: open the options menu in Copilot Chat, select Scheduled prompts, and choose Run now to execute immediately, Turn off to pause without deleting, or Delete to remove it entirely. "Pause without deleting" maps to Turn off — a clean distractor pair with Delete.

The licensing point is unambiguous and heavily tested: scheduled prompts require a Microsoft 365 Copilot license and run on work-grounded data. They are not available with Copilot Chat alone.

There's an admin dependency too. Scheduled prompts depend on connected experiences. If an admin sets the Allow the use of connected experiences in Office cloud policy to Disabled, users can no longer see the scheduled prompts icon or manage existing prompts — though previously scheduled prompts keep running until their schedule finishes. Scheduled prompts are supported across Microsoft Teams, Office.com chat, and Outlook.

Sharing happens from Prompt Lab or directly from a sent prompt. Select the Share prompt icon and you get two routes. Copy a direct link and paste it into email or chat — the right choice for one colleague who isn't on any of your teams. Or select a team, and everyone on that team sees the prompt under the Team prompts tab in Prompt Lab.

The distinction between those two routes is the exam question: a link suits individuals, sharing with a team suits everyone on that team. Watch for scenarios that name a single person outside your team, which rules out the team option.

Beyond you and your teammates, there's a third publisher: your organization's admin can publish prompt collections tailored to your company's workflows and terminology. These org prompts appear in Prompt Lab for everyone in the organization alongside Microsoft's Suggested prompts.

Your interactions are stored as Copilot activity history, which means you can browse past sessions or search them rather than starting from scratch. The flow is: open Copilot Chat in Microsoft 365, open your conversation history, use the search bar to enter keywords from a previous prompt or response, and select the conversation you want to revisit.

Searching works best with specific keywords — project names or distinctive task phrases rather than generic words. If you've been using agents, you can also filter your conversation history by agent to find and resume previous interactions with a particular one.

Scheduled prompts are marked in your conversation history, which makes their runs easy to pick out from ordinary chats.

For a single conversation: find the chat in your conversation history, select the options menu next to the chat title, select Delete, and confirm. For everything at once: sign in to the My Account portal, go to Copilot app activity history, and select Delete all activity history.

Several consequences are worth knowing because they make good distractors. Deleted conversations can't be restored. Deleting your history doesn't affect your organization's usage reports. Any shared conversation links remain accessible to others. And your organization's retention policies may affect what can actually be deleted — admins can set Purview retention policies for Copilot interactions.

One more that's easy to get wrong: deleting a conversation does not remove memories saved from that conversation. Saved memories are managed separately under Personalization settings.

Renaming is the simplest item in this domain, and it's tested mainly as a matter of knowing where the control lives: the options menu next to the chat title in conversation history, not in the chat itself.

The practical point behind it is organization. Copilot names conversations automatically based on the opening prompt, which produces titles that are hard to scan later. Renaming a chat to something like "Marketing Plan Draft" makes it far easier to locate by topic when you come back to it — which pairs naturally with searching your history.

Reach Notebooks from the Microsoft Copilot app by selecting Notebooks in the left navigation, or from OneNote by selecting Create Copilot Notebook. When you create a notebook you name it and add references — the files and sources Copilot grounds its answers in. From the References tab, Add reference offers suggested files based on your recent activity, search, upload from your device, attach cloud files from OneDrive, or a link to a supported file.

Supported reference types include Word, Excel, PowerPoint, PDFs, images, text and code files, SharePoint and OneDrive links, Teams chats, emails, OneNote text pages, and past Copilot Chat conversations. Only the first 100 references in a notebook are used for grounding — a specific number that makes for an easy exam question.

A notebook can carry its own standing instructions. Select Add Copilot instructions, enter your preferences for tone, formatting, language, or focus, and save. Those instructions apply inside that notebook only, which distinguishes them from custom instructions under Personalization, which apply to every conversation.

Sharing a notebook uses Share from the toolbar: enter names or email addresses, choose view or edit access, and send. Shared members can add references and ask Copilot questions grounded in the same content. Notebooks can also produce an audio overview, or generate a Word document or Excel spreadsheet from their contents via Create.

The distinction from Pages is a reliable exam question: a Notebook is a grounded workspace built around references you gather, while a Page is a shared, editable canvas built from a Copilot response.

The Agent Store is a curated marketplace reachable from the Copilot app. Microsoft-built agents include Writing Coach, Prompt Coach, and Visual Creator, each aimed at a specific productivity scenario. Browsing and installing from the store costs nothing to set up, which is why it's the default answer when a scenario describes a common, generic need.

Create a new agent when the work is specific to your team: your processes, your preferred knowledge sources, your tone and format. From the Agent Store, select Create an agent to open Agent Builder.

Two advanced agents live in the store but deserve separate attention because they're named in the audience profile. Researcher handles complex, multi-step research tasks, drawing on both the web and your work content — files, emails, meetings, and chats you have access to — and returns a comprehensive, source-cited report with organized sections and visuals. It deliberately takes longer in exchange for depth. Reach for standard Copilot Chat for a quick summary, short reply, or lightweight brainstorming; reach for Researcher when you need reasoning across multiple sources and a report you can share.

Analyst is the data counterpart. It analyzes data from Excel spreadsheets, CSVs, JSON files, PDFs, and Word documents, reasons across multiple files at once to identify trends and outliers, and produces structured, shareable reports with charts, tables, and cited findings. The three-way split is worth memorizing: Copilot in Excel for exploring and charting inside an existing spreadsheet, Analyst for deep analysis across multiple files with a polished report, and Copilot Chat for general chat, web research, and quick drafts. Both Researcher and Analyst require a Microsoft 365 Copilot license.

Agent Builder includes templates that apply design guidelines and industry best practices so you spend your time customizing logic rather than starting from an empty form. The steps: in Microsoft 365 Copilot, select New agent in the left pane; under Start with a template, select the template you want, and the agent is created automatically; then use natural language to update the name, description, and instructions, or work on the Configure tab to add an icon, knowledge sources, and capabilities.

There are three entry points when you choose New agent, and the exam can ask you to pick between them. Describe the agent in natural language and let Copilot build it — the recommended path. Skip to configure and set each field yourself manually. Or start with a template and customize from there.

Changes save automatically as you work. When the agent behaves the way you want, select Create in the upper right, then Go to agent to start chatting with it. It remains private to you until you choose to share it.

The specific limits are exam material because they're easy to write questions about. Up to four public website URLs, and those URLs can be at most two levels deep and cannot contain query parameters. Up to 100 SharePoint files, folders, or sites, plus at most one SharePoint list. Up to 50 OneDrive files. Up to five Teams chat URLs. Up to 20 files uploaded from your device. Copilot connectors if your admin has enabled them.

Permissions behave differently depending on the source, and this distinction is the most commonly tested point. SharePoint and OneDrive sources respect each user's existing permissions and sensitivity labels, so a colleague who can't access a folder simply gets answers without that folder's content. Files you upload from your device become embedded content, and anyone who can use the agent can see information from them — embedded files do not follow per-user file permissions.

Teams chats and meetings, Outlook email, and People data all require a Microsoft 365 Copilot add-on license. Email knowledge can't be scoped: adding it means the agent uses your whole mailbox, and users you share the agent with don't get access to your email.

The Only use specified sources toggle makes the agent prioritize your knowledge sources for search-based questions, falling back to a message when it can't find the information. Read the wording carefully: it prioritizes rather than fully blocks general AI knowledge, and Agent Builder cannot fully block it. For strict control, use Copilot Studio.

A practical note that shows up in troubleshooting scenarios: newly uploaded SharePoint or OneDrive files can take several minutes to become usable, and show as "Preparing" in the Knowledge section until they're ready.

Instructions tell the agent who it is, what it should do, and how it should respond — persona, tasks and goals, tone and format. For example: "Act as a financial analyst. Analyze quarterly sales data and respond in a clear, professional tone using bullet points." The character limit is 8,000. Keep this separate from the description, which is limited to 1,000 characters and exists to help the model identify when to use the agent at all. Name has a 30-character limit.

Capabilities are toggles that extend what the agent can do. Code interpreter — shown as Create documents, charts, and code — solves math problems, analyzes data, and generates visualizations. Image generator — shown as Create images — produces images from prompts. If a scenario mentions analyzing data or building a chart, that's code interpreter; if it mentions a picture or illustration, that's image generator.

Starter prompts, also called suggested prompts, are example prompts each with a name and description that show other users the scenarios the agent supports — "Summarize this report", "Create a project timeline". There's no minimum number. One consequence matters: if you edit suggested prompts after sharing the agent, people with access need to reinstall it to see the updated prompts.

The Try it tab is how you test an agent inside the authoring experience. It becomes available once the agent has a name, description, and instructions, and it behaves like the published agent. Some features aren't available there — you can't share prompts, provide feedback, or @mention other agents — but it's the right answer to "how do I test this before sharing it?"

You can also set a default response mode in the Model section: Auto (the default, balancing speed and depth), Quick response for straightforward questions, or Think deeper for complex tasks. Users can override this with the model selector.

Two roles exist in the share dialog. Can chat makes someone a chat user who can interact with the agent but not change it. Can edit makes them an owner with full rights to edit, share, and manage it — and all owners have equal rights, including adding and removing people. One limitation makes a clean exam question: you can't add a group as an owner. Groups can only be added as Can chat; to grant owner rights, add individuals.

Beyond named people, there are two broader options and they are not the same thing. The Org-wide sharing for chat access toggle makes the agent visible in the Agent Store to everyone in your tenant, so they can discover it and start chatting. Submitting to your org catalog is a separate flow where an admin reviews the submission and, if approved, publishes it in the Built by your org section of the Agent Store. Sharing with people gives direct access; the org-wide toggle gives discoverability; the org catalog gives an admin-approved listing.

Sharing an agent does not automatically grant access to its knowledge. The agent respects each user's permissions and sensitivity labels, so a user without access to a SharePoint source simply gets responses that don't draw on it. You can opt to share the specific SharePoint files and folders when you share the agent, but only if you already have permission to share them.

Updating a shared agent has a two-step rhythm that gets tested. Edits save automatically, but users don't see them until you select Update, and the change can take several minutes to reach end users. If the agent uses SharePoint file or folder knowledge, reshare it with the same users after updating so the files are shared again.

Finally, admins can restrict sharing at the tenant level — all users (the default), specific users or groups, or no users. Those changes apply only to new sharing actions and don't revoke existing access.

In Word, Copilot generates a first draft from a prompt and then lets you refine it by conversation: adjust the tone, change the length, restructure it, add a section. A prompt like "Write a professional introduction for our Q3 performance report. Audience: leadership team" produces something usable; "write an intro" doesn't. Follow-ups like "Make the tone more optimistic and forward-looking" continue shaping the same draft.

In PowerPoint, describing what you need produces a draft deck with structure, content, and slides — for example, "Create a five-slide leadership presentation on our Q3 performance highlights and key wins." This is the right path when you're starting fresh without a source document.

In Outlook, Copilot generates a complete email or meeting request from one prompt. Open a new message, select the Copilot icon in the toolbar, and describe what you need.

Whichever app you're in, the grounding rule from the fundamentals domain still applies: you steer the source through the prompt. Reference a file with / to draw on organizational content, or describe a web source to draw on the public web. Copilot can automatically include citations when it draws from work files or the web.

Before you share anything, Learn asks you to validate it against the business goal: does the tone and detail match the audience, does the content support the decision you need, is this the right format, and have you checked the facts and figures? If any of those fall short, refine the prompt or edit directly rather than sending as-is.

The mechanism is the / reference. In the Copilot pane, type / and start typing the file name; Copilot finds it and uses it as a source. "Create a five-slide leadership presentation based on /Q3_Performance_Report.docx" turns an existing Word document into a deck without copying anything by hand.

PowerPoint's five-file limit is a specific, testable number: you can reference up to five files at once and mix types — a Word document, an Excel file, and a PDF in one prompt.

Sensitivity label inheritance applies here and is worth carrying over from the fundamentals domain. When Copilot in Word, PowerPoint, or Outlook creates new content from a labeled item, the new content inherits the source's label and protection settings. If several labeled files were used, the highest-priority label wins.

Copilot can also help you understand a document before you change it. Switch to Chat only mode so Copilot answers in the pane without editing the document, then ask for a summary, for the decisions still needed, or for the unresolved risks. That's the right approach for an inherited document or a long report you need to get up to speed on.

A management summary is a summary with constraints. "Summarize this document" gives you a summary; "Summarize this document in five bullet points for the leadership team, focusing on decisions needed" gives you something a manager can act on. The four prompt ingredients apply directly — the audience is the context, the bullet count is the expectation.

In Word, switch to Chat only mode and ask for the summary so Copilot responds in the pane without modifying the document. Useful prompts include "Summarize this document in five bullet points", "What decisions are needed before this document is ready to share?", and "Identify the open questions or unresolved risks in this document."

The same content can be reshaped for different audiences without redoing the analysis. With the Analyst agent, for example, follow-ups like "Summarize these findings into three bullet points for a leadership slide", "Write an executive summary for this report", or "Draft an email update summarizing these insights for the team" reformat an existing report rather than re-running it.

Verification matters more here than almost anywhere else, because a management summary is what a decision gets made on. Check the facts, figures, and citations before it leaves your hands.

From chat to a Page: select ••• underneath a Copilot response and choose Edit in Pages. The content copies into a new editable page you can keep building on with Copilot, side by side with the chat. Use Add to Copilot Page to append a later response to a page you already started.

From a Page outward: in the upper-right corner of the page, the Word icon converts it into a Word document and the PowerPoint icon turns it into a starting slide deck. Both produce standalone files you can share with people who don't have access to the original page.

From a meeting to a document: open the Recap tab in the meeting chat and use the AI summary, which can be exported to Word or Excel. To draft a follow-up, use Copilot Chat and reference the meeting with / followed by its name, which grounds the response in what was actually said — for example, "Draft a follow-up email for the LaunchPad Team summarizing /LaunchPad Project Check-In."

Between apps generally: reference the source with / and name the destination in the prompt. Excel data into a PowerPoint deck, web data into Excel, analysis results into an email — the pattern is the same each time. Copilot Notebooks can also generate a Word document or Excel spreadsheet from their references via Create.

Where these questions get interesting is when two routes both look plausible. Read for the end state: if the result needs to be collaborative and living, that's a Page; if it needs to be a file someone can be sent, that's a Word or PowerPoint export.

Before the meeting, Copilot can prepare you. From the Copilot pane in Outlook or from Copilot Chat, a prompt like "Summarize everything related to the LaunchPad project from my emails, files, and meetings over the past two weeks" pulls the relevant material together in one step. Copilot in Outlook can also create the invite itself — switch from Default mode to Allow actions so it can act on your calendar, then review the proposed title, attendees, and agenda and select Approve.

During the meeting, Copilot summarizes discussion in real time including who said what and where people agree or disagree, suggests action items and follow-ups, answers questions from the transcript and chat, and builds tables and lists from meeting content. The prerequisite is the thing to remember: start recording first, because that generates the transcript Copilot needs. If recording isn't available, use Transcribe. Without a transcript, Copilot is limited to chat-based data only.

Facilitator is a distinct agent and the contrast with Copilot is the exam point. Your interactions with Copilot in a meeting are private to you; Facilitator is shared — when anyone types @Facilitator in the meeting chat, all participants see the interaction. Facilitator automatically takes collaborative notes, tracks topics with visual timeline markers, and can generate a Word document of notes and action items afterwards. It must be enabled by an IT administrator.

After the meeting, intelligent recap is the named feature. Open the Recap tab in the meeting chat and select AI summary for a structured summary of what was discussed, who contributed, where the team aligned or disagreed, and what follow-ups were proposed. It's available for scheduled, instant, and ad-hoc meetings when a transcript exists. You can browse the recording by speaker and topic, and summaries include citations and export to Word or Excel.

Create a page from any response you want to build on — select ••• underneath the response and choose Edit in Pages, or ask Copilot directly to "Create a page about…". The page opens side by side with your chat so you can keep prompting and adding to it. Use Add to Copilot Page to append later responses to a page you've already started.

Everything on the page is directly editable, and the / key opens a menu of content types: tables, headings, checklists, code blocks, and more. Pages are saved automatically and found from the Pages section in the Copilot app, the Pages tab in Microsoft 365 on the web, or the link in the originating chat.

Sharing works by @mentioning people in the page, after which they can edit content in real time. The key distinction: people you share with get access to the page and its content, not to your Copilot Chat session. Pages tracks every edit — select any text to see the contributor's picture, or a lead contributor followed by +N when several people edited a section. You can also share a direct link or embed the page as an interactive component in Teams or Outlook.

Retention has a wrinkle worth knowing: pages you create are saved indefinitely unless you delete them, but if you're not the owner and that person leaves the organization, their pages are deleted after 30 days unless you make a copy.

Pages versus Notebooks is the reliable comparison question. A Page is a shared, editable canvas that starts from a Copilot response and is for producing and collaborating on content. A Notebook is a grounded workspace that starts from references you gather and is for asking questions across that material.

Custom instructions are a one-time setup that tells Copilot how to respond across all conversations — your role, your communication preferences, recurring context you'd otherwise repeat every time. For example: "I'm a project manager. Keep responses concise and use bullet points. Avoid technical jargon." Set them in Copilot Chat settings under Personalization, then Custom instructions. The rule of thumb is to move whatever you repeat in every prompt into custom instructions so your prompts stay focused on the task.

Copilot Memory is the other half. Copilot remembers relevant details from your conversations and applies them later, building memory two ways: saved memories, which Copilot may prompt you to save or which you create by saying "Remember that I prefer Excel over PowerPoint for reporting"; and chat history inferences, which are patterns it recognizes from previous conversations. Memory is available to both Copilot Chat and licensed Microsoft Copilot users, and admins control whether it's available at all.

Managing memory lives under Personalization, where you can view saved memories, delete a specific memory, turn off chat history so past conversations don't personalize future responses, and update or clear custom instructions. Three rules there are exactly the kind of detail the exam tests: deleting a conversation doesn't remove memories saved from it; turning off saved memories stops new ones but doesn't delete existing ones; and Delete all memories is what actually clears what Copilot has stored.

Temporary chat is the escape hatch. In a temporary chat, Copilot doesn't access your saved memories or custom instructions, the conversation doesn't appear in your chat history and can't be referenced later, and no memories are saved from the session. It's the right answer whenever a scenario asks for a fresh, neutral response uninfluenced by past context. Note that temporary chat data may still be retained under your organization's retention policy.

Don't confuse these with notebook instructions, which apply only inside one Copilot Notebook, or with agent instructions, which shape one agent's behavior. Custom instructions apply to all your conversations.

Microsoft frames three ways of working and the difference is your role in each. Copilot Chat is a conversation — you're in it, steering every turn. Agents are purpose-built helpers for repeatable, scoped tasks in a defined area. Cowork is an assignment — you describe the outcome, hand it off, and Copilot runs with it across your apps, files, meetings, and calendar while you move on to something else.

Cowork suits multi-task work: preparing everything for a customer meeting, building a competitive briefing with research and a shareable summary, coordinating follow-up from a meeting including action items and draft messages. With Chat you'd manage those one prompt at a time; with Cowork you hand off the goal.

A useful signal: if your prompt starts to read like a list of steps — first do this, then that, then create — you're describing a larger goal, and that's a sign to consider Cowork instead of managing every step yourself.

Cowork may fit when the work spans email, files, meetings, calendar, or Teams; when the outcome needs more than one deliverable; when you'd rather come back to results than manage each step; or when the work continues over time or runs on a recurring schedule. It appears in Microsoft Copilot on web, desktop, and mobile.

Once you hand off a goal, Cowork plans the steps, manages files and tasks across apps, and delivers completed work, checking in with you along the way. The interaction pattern is deliberately different from chat: you step away rather than staying in the loop for each turn.

Skills are how Cowork specializes. Alongside the built-in ones, you or your organization can create custom skills, and plugin skills come from third-party publishers. Both carry a caveat Microsoft states plainly: custom skills aren't validated by Microsoft, so their quality depends on how they were written, and plugin skills' quality depends on the publisher. Review the output of either carefully.

Web search applies to Cowork as it does to Researcher and Analyst. It isn't a prerequisite, but enabling it is recommended to get the most value. Note that Cowork has no user-facing Web search toggle, and that if an admin chooses the option to disable web search in Copilot Work mode while leaving it on elsewhere, web search in Cowork is disabled.

The security model is identical to the rest of Copilot, which is the main point to take away. Cowork acts as you using your Microsoft Entra ID sign-in and your access scope — if you can't see a file, neither can Cowork. Sensitivity labels and DLP policies apply to all Cowork interactions. Your data stays isolated to your tenant, and Cowork doesn't use your organizational data to train models.

Microsoft's own three-point summary is the cleanest way to hold it: Cowork sees what you see; Cowork doesn't take sensitive actions without you, so sending, sharing, and scheduling wait for your approval; and Cowork doesn't train on your data.

The stated limitations are specific. Vague or overly broad instructions can be misinterpreted. AI-generated documents, emails, and messages are drafts to review before you approve sending or sharing. Cowork can't access files on your local device — it works with OneDrive, SharePoint, and other connected cloud services. It can't delete files or folders. Results may be incomplete if source data isn't indexed or is out of date. Complex multi-step tasks with many dependencies may not always finish as expected. And voice input support varies by browser.

Microsoft is explicit that Cowork isn't intended for use cases requiring guaranteed accuracy without human review — legal filings, medical decisions, or financial transactions that bypass approval processes.